[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

Re: Gnu Questions..

Posted By: Daniel Murphy
Date: Wednesday, 18 April 2007, at 10:28 a.m.

In Response To: Gnu Questions.. (geoff arnold)

What settings do you use for rollouts with gnu?

My default rollouts settings that I use for 95% of my rollouts are:

  • Trials -- depends on how close the plays are and how accurate I want the equities to be. You can judge from the evaluations how many trials might be needed. I run a minimum of 108 trials and usually 324, 648, 1296 or more, but I would emphasize that you do not need to run 1296 trials of every rollout and it's really not necessary that the number of trials is divisible by 36. In a rollout I'm looking for a JSD of 3. Once I get that, more trials are extremely unlikely to change the decision and will only make the equities more accurate.
  • Truncation OFF
  • Evaluation for later plies OFF
  • Stop when result is accurate OFF
  • Stop rollouts of multiple ... OFF
  • Cubeful ON
  • Variance reduction ON
  • Quasi-random dice ON
  • Rollout as initial position OFF
  • Cube decisions same settings OFF
  • Use same settings for both players ON
  • Use player0 setting for truncation point ON
  • Checker play: 2-ply
  • Cube: 2-ply
  • Use neural net pruning ON
  • Cubeful checker evaluation ON
  • Move filter NARROW
  • MET = G11

Albert's suggestions at AllAboutGNU are reasonable. I agree that "0-ply checker play is indeed extremely strong" and since it is fast it gets results in good time but since I usually weigh accuracy more heavily than speed I do not agree that 0-ply is "usually ideal for full rollouts." I don't use 0-ply checker play except for simple noncontact positions and never use 0-ply cube. Even in the simplest of contact positions, like a bearoff against checkers on the roof, 0-ply and 2-ply GWCs may differ by 1% or more. Albert's "third setting" recommendations are o.k. The Ians' rollouts with "separate evaluations" enabled changing to 0-ply checker play after 12 plies have usually gotten good results compared to full 2-ply rollouts but if the position is likely to be complex after 12 plies then a full 2-ply rollout would be a better choice. I have found that there is very little difference in results between 2-ply with Narrow, Large (world class) and Huge (supremo) move filters.

Messages In This Thread

 

Post Response
Your Name:
Your E-Mail Address:
Subject:
Message:

If necessary, enter your password below:

Password:
 

 

[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums is maintained by Stick with WebBBS 5.12.